Meta is making a clear statement about where the next phase of AI infrastructure is heading. In an expanded partnership with Amazon Web Services, the company has committed to using tens of millions of Graviton5 CPU cores – positioning itself as one of the largest customers of Amazon’s custom ARM-based processors.
At first glance, this might look like just another large-scale cloud deal. But the underlying shift is far more strategic.
For years, the AI conversation has been dominated by GPUs. Training large language models, running inference at scale, and competing in the generative AI race have all been tied to massive GPU clusters. However, Meta’s latest move highlights a different layer of the stack – one that is becoming increasingly critical: Agentic AI.
Unlike traditional models, agent-based systems are not just generating outputs. They plan, orchestrate, execute and iterate across complex workflows. This requires a different type of compute architecture. While GPUs remain essential for training, CPUs are becoming the backbone of execution, coordination and real-time decision-making.
This is where Amazon’s Graviton architecture comes into play. The Graviton5 generation, introduced at re:Invent 2025, is designed for high-performance, energy-efficient cloud workloads, offering 192 cores and a significantly expanded cache. Built on ARM architecture, it reflects a broader industry trend away from traditional x86 dominance.
Meta’s decision is not happening in isolation. Competitors are moving in the same direction. Google is pushing its Axion chips, while Microsoft is investing in its Cobalt processors. At the same time, ARM itself has stepped deeper into the AI space, recently unveiling its own AGI-oriented CPU concept in collaboration with Meta and TSMC.
From a Darkgate perspective, the implications go beyond infrastructure optimization. What we are seeing is the emergence of a dual-layer AI economy:
On one side, GPU-heavy training environments define who can build models. On the other side, CPU-driven execution layers define who can operate, scale and monetize AI systems in real-world environments.
The shift toward agentic systems also introduces new challenges. More autonomous systems mean more distributed decision-making, more API interactions, and a significantly larger attack surface. The infrastructure powering these agents – often invisible in traditional security models – becomes a critical point of control.
Meta’s investment in Graviton is therefore not just about cost efficiency or performance. It is about positioning itself for a future where AI is no longer a static model, but a network of autonomous systems operating continuously in the background.And in that world, compute is no longer just about power. It’s about orchestration.
